A Method for Single Image Reflection Removal Based on Binocular Vision and Nonlinear Activation Modules

Ensuring cabin visual safety is critical for autonomous driving systems, where lens reflections and sunlight glare often compromise perception reliability. Single image reflection removal (SIRR) has long been a challenging task in computer vision due to its ill-posed nature. However, removing reflections from misaligned image pairs is even more underexplored and difficult, as precise correspondence is absent. In this paper, we observe that binocular stereo images captured from dual cameras in autonomous driving scenarios are naturally spatially misaligned, and often, only one view may contain reflections. This creates an ideal setting for training with misaligned reflection pairs while still supporting single-image inference. Leveraging this property, we propose a deep learning-based method for reflection removal in misaligned stereo images, aimed at enhancing cabin perception reliability. Our model is first trained on standard aligned datasets, then fine-tuned on a real-world misaligned dataset of 9,456 image pairs collected using the Zhongke Huiyan binocular camera. A novel nonlinear activation module is introduced to improve adaptation to illumination variation and structural consistency. Our method performs effectively on both aligned and misaligned data, requiring only a single contaminated image at inference. Experiments show our approach reduces the LPIPS score by 8.76 % over raw stereo input and surpasses the state-of-the-art ERRNet by 1.1 % on misaligned image benchmarks. These results demonstrate the potential of leveraging stereo misalignment for reflection removal, providing a practical solution for enhancing cabin visual safety in real-world autonomous driving environments.